Sizing Lakebase Search: Lessons from a Real App and 4 Million Passages
Databricks, Thursday, October 15th, 2026
Databricks shares benchmarks and sizing guidance for Lakebase Search hybrid vector and keyword search in Postgres.
Lakebase Search runs vector and BM25 keyword search as Postgres extensions inside the application database, avoiding a separate search system, and Databricks benchmarked it using an internal app and up to 4 million passages.
Sizing the local file cache to match the vector index kept 95-98% of pages in memory and cut warm vector latency to single-digit milliseconds.
A 14-minute load test serving 109,000 requests used just $0.02 of compute, and Databricks recommends a 0.5 CU always-on floor for production.